YOLOv4

Understand YOLOv4's architecture, its Bag of Freebies and Specials, and why it is a fast, reliable real-time object detector.

5 parts, in reading order · 6 interactive figures

Start with part 1

  1. YOLOv4 — Version 0: Introduction

    Part 1 · Paper · May 2020 · 1 interactive figure

    • YOLOv4 uses CSPDarknet53 backbone, SPP and modified PANet neck, and YOLOv3 head for its architecture.
    • Three backbone candidates were compared on parameters, receptive field, FLOPs, and FPS on RTX 2070.
    • Bag of Freebies adds training augmentations like Mosaic, CutMix, DropBlock, and CIoU loss without inference cost.
    • Bag of Specials adds Mish activation, CSP connections, and MiWRC for a small inference cost increase.

  2. YOLOv4 — Version 1: Bag of Freebies

    Part 2 · Paper · Jun 2020 · by Deval Shah on VisionWizard ↗

    Yolov4 is real time object detection algorithm. It is successor of yolov2 and yolov3.

  3. YOLOv4 — Version 2: Bag of Specials

    Part 3 · Paper · May 2020 · 5 interactive figures

    • Mish activation replaces ReLU, avoids dead neurons with a smooth, non-monotonic curve that dips slightly below zero.
    • Mish + CSPDarknet53 backbone raises AP from 42.4 to 43.0 on COCO 512x512 in the paper's ablation study.
    • CSP (Cross Stage Partial) connections split feature maps to reduce DenseNet's heavy concatenation cost while keeping gradient flow.
    • CSPDarknet53 with Mish and Bag-of-Freebies backbone achieves 43.0 AP, 64.9 AP50, 46.5 AP75 at 512x512 resolution.

  4. YOLOv4 — Version 3: Proposed Workflow

    Part 4 · Paper · Jun 2020 · by Deval Shah on VisionWizard ↗

    The architecture of yolov4. Fastest and most reliable object detection.

  5. YOLOv4 — Version 4: Final Verdict

    Part 5 · Paper · May 2020

    • Bag of Freebies ablation shows S+M+IT+GA with CIoU loss reaches 42.4% AP on CSPResNeXt50-PANet-SPP.
    • Adding SAM attention to CSPResNeXt50-PANet-SPP improves AP to 42.7% over baseline 42.4%.
    • Classifier ablation on CSPResNeXt50 finds CutMix+Mosaic+Label Smoothing+Mish yields 79.8% Top-1 accuracy.
    • Classifier ablation on CSPDarkNet53 shows CutMix+Mosaic+Label Smoothing+Mish reaches 78.7% Top-1 accuracy.

Other series